Short-term breast cancer risk prediction improves when AI combines imaging with personal info
Integrating clinical and mammographic information into risk assessment models can better identify which women face the highest short-term risk of developing breast cancer.
Understanding which women have the greatest short-term risk could enable providers to implement targeted screening strategies to ensure malignancies are caught at the earliest possible stage. In turn, patients could seek less invasive treatments sooner, before disease has progressed.
There are multiple studies that incorporate imaging and clinical data to assess patients’ long-term risk, but information regarding the utility of these models for short-term risk is lacking. The authors of a new analysis in European Radiology sought to address this by developing predictions models that assess patients’ short-term (2 years) risks.
“Long-term risk models (5–10 years) are valuable for developing population-based approaches as well as designing screening strategies. However, AI-based models have shown potential for short-term prediction, allowing for targeted assessments of women at high risk. This aligns with the standard screening interval in Europe, which typically ranges from two to three years,” Rafael Llobet, from the Institute of Information Technology, Polytechnic University of Valencia, and colleagues noted. “Unlike long-term models, which primarily inform general screening policies, short-term models are particularly useful in identifying high-risk individuals who may benefit from earlier intervention through more frequent monitoring or additional diagnostic procedures.”
The team developed prediction models using data from more than 2,000 women who underwent breast cancer screening at a single institution between 2013 and 2020. The cases included 418 women whose mammograms were taken up to two years before they were diagnosed with breast cancer and another 1,775 who stayed cancer-free for at least two years.
In total, three models were created. One combined personal information and mammogram images using a method called “Extremely Randomized Trees” (ERTpd + im); one incorporated images only, using a convolutional neural network (CNN), while a third hybrid model combined both methods.
Of the three, the hybrid model yielded the best performance, achieving an AUC of 0.75; in comparison, the CNN model and the ERTpd + im model achieved AUCs of 0.74, respectively. The hybrid model’s performance was consistent across breast density categories. It also was more effective at predicting screen-detected cancers than interval cancers, though this finding could be owed to the relatively few instances of interval cancers included in the training data.
“Our study confirms that images and clinical data contain information with predictive capabilities, suggesting that AI models use mammograms as the main source for prediction, but also obtain complementary information from risk factors and other additional characteristics,” the authors explained. “This contribution could justify the superior predictive power of AI compared to traditional statistical procedures when assessing breast cancer risk, since they do not consider the images themselves.”
Read more about the research here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.